Anthropic spend per token on OpenRouter is 2.1x that of OpenAI

The shift in spend dynamics

Recent data from OpenRouter reveals a striking divergence in how users are spending money on the two leading AI labs. While the volume of tokens produced by Anthropic and OpenAI is nearly balanced, the actual cost to the user is not. In the most recent week of data, Anthropic captured 48% of the tokens but commanded 66% of the total spend.

Conversely, OpenAI held a slightly larger share of the token volume at 52% but only received 34% of the spend. This means that for every token used, Anthropic's models are costing users about 2.1 times more than OpenAI's models. This disparity suggests that users are opting for Anthropic's models for specific high-value tasks that justify a higher price point.

As reported by FourWeekMBA, these figures come from a chart shared by Peter Walker, the Head of Insights at OpenRouter. The data highlights a clear trend where Anthropic is capturing a disproportionate amount of revenue relative to its actual usage volume. This suggests a shift in the market toward premium, high-cost outputs over high-volume, low-cost ones.

Model performance and pricing

The data points toward a specific turning point in the competitive landscape between the two companies. Peter Walker noted that the introduction of Opus 5.5 served as a catalyst for this shift. Before this, OpenAI had been making steady gains in spend share throughout the year, but the new Anthropic model appears to have reversed that momentum.

The pricing dynamics are complicated by various factors, including how different models are bundled and the underlying hardware costs. For instance, OpenAI has mentioned plans to share usage allowances between ChatGPT and Codex starting in December. Meanwhile, hardware costs remain a significant variable, with some data showing Nvidia's B200 costing up to 2.3 times more than AMD's MI355X per token for certain models.

When looking at the broader market, the cost of production is actually trending downward. Since the first week of 2026, total tokens on OpenRouter have increased 224 times, while total spend has only increased 81 times. This indicates that the overall spend per token has fallen by about 64% since the beginning of the year, even as Anthropic maintains a premium position.

The economics of tokens

To understand these numbers, it helps to look at the concept of tokenomics, which describes the token as the fundamental unit of AI economics. A token is simultaneously the unit of cognition produced by the model, the unit of compute served by the data center, and the unit of price charged by the lab. Because it serves all these roles, the relationship between volume and spend is the primary lever for business strategy.

In this framework, a lab's revenue on a router like OpenRouter depends on the balance between how many tokens it serves and the price of each one. The current data shows that Anthropic is successfully positioning its models as high-value units of cognition. Users are choosing to spend more on fewer tokens from Anthropic than they are spending on a higher volume of tokens from OpenAI.

This behavior suggests that for many developers and enterprises, the value extracted from the AI is currently tied more closely to the quality of the output than the quantity. While OpenAI's models may be more cost-effective for high-volume tasks, Anthropic's models are winning the spend share for tasks where precision or specific capabilities are prioritized. This creates a bifurcated market where different models serve different economic needs.

Market limitations and context

While these figures provide a clear snapshot of the competition, it is important to view them within the context of the OpenRouter platform. OpenRouter is just one of many ways to access these models, so its data reflects the preferences of its specific user base rather than the entire global market. The chart specifically compares Anthropic and OpenAI against each other, meaning the percentages represent their combined share of the platform's traffic.

Furthermore, the data does not provide absolute numbers for total tokens or total spend. It also does not specify the exact amount of spend that reaches each lab after OpenRouter takes its own fees. Therefore, while the 2.1x spend ratio is a significant finding, it describes the user-facing economics on this specific platform rather than the internal margins of the AI labs.

Looking ahead, the trend suggests that the